Project Summary

• Contributed to the development of sarpyx, the ESA Φ-lab Python toolkit for advanced Synthetic Aperture Radar (SAR) processing, by developing and integrating Sentinel-1 (S1) sub-aperture decomposition processing.
• Defined a quantitative validation framework to select robust sub-aperture configurations for the generation of sub-aperture-derived features used in the WORLDSAR dataset construction workflow.
• Developed a self-supervised enhancement approach for S1 Stripmap imagery, exploiting full-aperture/sub-aperture pairs generated from the same acquisition, and disseminated the work through a paper at the AI4Space Workshop of CVPR 2026.
• Produced reusable processing concepts, validation criteria, and experimental material to support reproducible SAR processing and AI-ready SAR dataset generation.

Development Tools

Development Outputs

• sarpyx contribution: Development and integration of Sentinel-1 sub-aperture decomposition processing in the ESA Φ-lab SAR processing toolkit. Repository: https://github.com/ESA-PhiLab/sarpyx
• WORLDSAR validation framework: Definition of the criteria and analysis workflow used to select the sub-aperture configuration for sub-aperture-derived feature generation.
• Self-supervised Sentinel-1 enhancement model: Experimental model for enhancing Sentinel-1 Stripmap imagery using full-aperture/sub-aperture pairs generated from the same acquisition.

Project Description

SAR is a fundamental Earth Observation technology because it can acquire information about the Earth's surface independently of cloud cover and solar illumination. The Copernicus S1 mission is particularly relevant for operational monitoring and machine learning applications because it provides systematic C-band SAR observations over large areas and under a wide range of environmental conditions. Despite this potential, SAR data remain challenging to integrate into Artificial Intelligence (AI) workflows. SAR images are affected by speckle, acquisition-mode dependencies, complex-valued signal properties, and specific geometric processing requirements. These characteristics make it necessary to design processing pipelines that are both physically meaningful and reproducible before SAR data can be used at scale in AI-ready datasets.

The collaboration with ESA Φ-lab focused on this interface between SAR signal processing, open-source software development, and AI-based Earth Observation. The work was organized around three connected activities. First, the project contributed to the development of sarpyx, with an emphasis on integrating S1 sub-aperture decomposition processing. Second, it defined a validation framework to select suitable sub-aperture configurations for generating features used in constructing the WORLDSAR dataset. Third, it developed a self-supervised enhancement model for S1 Stripmap images by exploiting full-aperture/sub-aperture image pairs generated from the same SAR acquisition. Together, these activities aimed to improve how S1 data can be transformed into richer, more informative, and more reproducible inputs for downstream AI models.

A first outcome of the collaboration was the contribution to sarpyx, an open-source Python tool developed in the ESA Φ-lab ecosystem for advanced SAR processing. The work focused on the development and integration of sub-aperture decomposition processing for S1 imagery. In SAR, the full synthetic aperture can be divided into different portions of the azimuth spectrum, producing several sub-aperture looks of the same scene. These looks preserve the same acquisition geometry while providing complementary observations that can reveal angular, structural, and radiometric differences. Sub-aperture processing is therefore useful not only as a signal-processing technique but also as a means of generating additional representations for AI workflows.

The implementation work required attention to both algorithmic and engineering aspects. From the SAR-processing perspective, the decomposition had to produce meaningful sub-aperture products that remain consistent with the original S1 acquisition. From the software perspective, the generated products had to be usable in downstream workflows, including SNAP-based processing, raster inspection, feature extraction, and machine learning experiments. This involved handling complex-valued SAR components, polarisation channels, product naming conventions, metadata consistency, and compatibility with later processing steps such as calibration and geocoding. The outcome is a reusable processing capability that helps bridge the gap between experimental SAR signal processing and operational-scale data preparation.

The second main component of the project addressed the problem of configuration selection for sub-aperture-derived features. Sub-aperture decomposition involves processing choices, such as the number of looks and the way the aperture is split. These choices affect the resulting products and the derived features. A configuration may generate visually distinct outputs while still producing highly redundant features, or it may increase diversity at the expense of stability. For this reason, configuration selection should not rely only on visual inspection. It requires a quantitative framework able to compare alternatives in terms of the information they provide and the robustness of that information across different samples.

Within the collaboration, a validation framework was defined to support the configuration used for creating sub-aperture-derived features in the WORLDSAR dataset construction workflow. The framework compared candidate configurations using criteria related to feature complementarity and stability. The objective was to identify a configuration that provides useful additional information while avoiding excessive duplication between features. This is particularly important for large-scale SAR datasets, where the selected representation should be informative enough for downstream learning but also efficient and reproducible across multiple products and scenes.

The validation considered different types of sub-aperture-derived information, including features related to covariance and coherence. Evaluating the behavior of these features separately is important because each feature family can respond differently to the choice of sub-aperture configuration. For example, the stability of diagonal covariance terms may differ from the behavior of cross-covariance or coherence-derived features. The framework, therefore, encouraged a balanced interpretation of the metrics, rather than selecting a configuration based on a single indicator. This approach provides a more defensible basis for dataset construction and supports future comparisons when new configurations or acquisition scenarios are considered.

A third outcome of the collaboration was the development of a self-supervised enhancement model for S1 Stripmap imagery. SAR image enhancement is relevant for many downstream applications, including infrastructure monitoring, urban analysis, object detection, and semantic segmentation. However, supervised SAR enhancement is difficult because clean, high-resolution SAR ground truth is typically unavailable. The project addressed this limitation by exploiting a physically consistent relationship within the S1 acquisition itself: sub-aperture images can be treated as degraded observations, while the corresponding full-aperture image serves as a higher-quality reference.

This strategy enables the construction of training pairs directly from SAR data without relying on external optical imagery, simulated degradation, or temporally separated acquisitions. Because the full-aperture and sub-aperture images are generated from the same acquisition, they are naturally aligned and represent the same scene at the same time. This reduces the risk of introducing inconsistencies caused by temporal changes or cross-sensor differences. The resulting learning setup is self-supervised in that the supervisory signal is derived from the internal structure of the SAR acquisition rather than from manually labeled or externally acquired reference data.

The enhancement model was designed to improve the visual and analytical quality of S1 Stripmap images while preserving relevant SAR structures. The task combines elements of speckle reduction and spatial-detail enhancement. This balance is important because excessive smoothing may remove meaningful scattering patterns, while excessive sharpening may introduce artifacts that are not physically supported by the data. The model development, therefore, focused on producing enhanced images that remain consistent with SAR-image characteristics and are suitable for subsequent Earth Observation analysis.

The scientific outcome of this activity was the preparation and publication of a paper at the AI4Space Workshop of CVPR 2026. This dissemination step was an important outcome of the collaboration, as it transformed the technical work conducted during the visiting researcher period into a scientific contribution to the AI and space communities. The work also demonstrates how SAR-specific signal processing can be used to design self-supervised learning strategies that do not depend on conventional labeled datasets.

Overall, the collaboration produced outcomes at three complementary levels. At the software level, it contributed to the sarpyx tool by integrating S1 sub-aperture decomposition processing. At the dataset level, it provided a validation framework to support the selection of sub-aperture configurations for WORLDSAR feature generation. At the scientific level, it led to a self-supervised S1 enhancement approach and a publication at an international AI-for-space conference. These outcomes contribute to the broader objective of making SAR data more accessible, reproducible, and useful for AI-based Earth Observation applications.

The following table summarises the main outcomes of the collaboration.

Figure 1: Sub-aperture decomposition simplified schema.

Figure 2: General workflow of the S1 Enhancement methodology (sub-aperture generation, training, and inference).

Figure 3: Visual comparison (VV polarization) between the MERLIN baseline and the proposed SI and MF methods.

Figure 4: Visual results on real full-aperture S1 SM imagery (VV polarization), comparing the proposed MF1 versus MERLINFull-aperture